Competitor compute and growth dockets#

Status: Reference Last verified: 2026-10-04 Canonical for: competitor docket coverage, comparison method and research queue; individual findings remain dated snapshots.

Scope#

Turn public infrastructure disclosures and product/growth evidence into comparable competitor dockets. The question is not just who owns GPUs: it is how capacity procurement, inference engineering and distribution support affordable acquisition, retention and paid usage.

This inventory covers the repository notes inspected, not every competitor ever discussed. It does not change our ownership goal, fleet purchases, pricing or scheduling policy. Those belong to the owned-GPU strategy.

Model#

Each docket records the operator, procurement evidence, acquisition offer, growth channels, monetization, dated outcomes, inference and unanswered questions. A disclosed deployment is positive evidence at that date—not a complete present-day fleet inventory.

The rows summarize the cited evidence in each linked docket. Every listed consumer/hosting competitor has a dated public-source pass. Some live offers remain inaccessible or contradictory; these limits are documented rather than replaced by old promotional numbers. Procurement contracts and conversion cohorts remain unavailable across most competitors.

Adjacent inference/API lane#

These providers compete for developer workloads or supply applications. They are separate from the consumer/hosting comparison above.

How it works#

The same questions guide each investigation:

1. Who operates the serving stack, and what proves purchased hardware, leases, cloud instances or external model services?
2. What is already committed versus billed per use? What inference/cache/scheduling optimization is disclosed?
3. What can users get free or cheaply, under which limits? Are credits a trial, creator marketing budget or recurring relaxed capacity?
4. Which acquisition channels and repeated-use needs are actually documented?
5. How does usage become revenue? Separate subscriptions, usage packs, APIs and team products.
6. What growth result is reported, by whom and when? Distinguish registered users, MAU, traffic, ARR and realized revenue.
7. What follows for ComfyICU, and what evidence would test that inference?

The working conclusion is narrower than “competitors grew because they owned GPUs.” Low-cost controlled capacity can fund acquisition, but OpenArt also documents rented infrastructure and deliberate distribution investment. No docket establishes a causal hardware-ownership → market-share → revenue chain.

2026-10-04 synthesis#

The fixed inventory now has 22 dated dockets: 19 consumer/hosting entries and 3 adjacent API providers. These are product research entries, not 22 independent companies: ThinkDiffusion/Floyo share history, and other company ownership relationships are recorded in their dockets. Public-source coverage is complete for this pass; successful Comfy worker execution and physical placement are not verified.

What the hardware evidence supports#

Other dockets leave serving procurement unknown. Do not turn a named GPU, a CDN/storage dependency, a private-node feature or a customer on-premises option into an ownership claim. A supplier inference is labeled as such.

What the growth evidence supports#

Cheap capacity is an enabling input, not an acquisition channel by itself. The dockets document search/creator/affiliate distribution, upstream open-source audiences, popular models, educational workflows, reusable projects and developer integration. None provides enough controlled cohort data to isolate free compute's contribution to retention or revenue.

Analyst implication: our owned-capacity strategy can improve control and fund bounded gap-fill acquisition, but needs a distribution surface and a repeat-use workflow. This research does not change our existing paid-work protection, short preemptible free lane or no-rented-spill rule. It does not authorize purchases, grants, pricing changes or product implementation.

Comfy pricing and execution conclusion#

Use Runpod Pod rental $2.09/hour versus Comfy API $4.54/hour as the main RTX PRO 6000 comparison requested by the founder: a $2.45/hour difference, 117.2% premium and 2.17× retail rate. The compute-only crossover for an always-on Pod is 46.0% Comfy billed-worker time in the same period. Storage, startup/idle, operations, availability and managed features remain unmatched. The separate approximately 30% Serverless rate pattern and US-NE-1 catalog correspondence support Runpod integration, not supplier costs or hardware ownership. Detailed comparison and primary citations.

The existing key exposed read-only catalog and billing data, including positive funds, but reports FREE. A tiny regular Cloud workflow returned HTTP 403 FREE_TIER_NOT_ALLOWED; private Build creation returned HTTP 402 despite funds. Thus lack of execution is an entitlement limit, not insufficient credits. No GPU worker ran. Exact worker placement requires different existing paid-plan access; no upgrade or purchase was authorized. Callback services were stopped, and deletion of the uploaded 1.7 KB probe blob remains unconfirmed. These limits close the current diagnostic pass without claiming a successful deployment or IP measurement.

Invariants#

• Self-operation, hardware title, reserved capacity and model-API integration are different claims.
• Unknown contracts/utilization prevent numerical claims about our cost advantage. Advertised retail prices are not suppliers' costs.
• Company and vendor outcomes are labeled reported, not independently audited. Read date never replaces publication date.
• Our ownership goal remains fixed. Competitor research does not authorize changes to the protected paid lane, short preemptible gap-fill lane or rented bridge.
• Dockets are append-only dated records. Update this index when coverage changes; append a new snapshot rather than silently rewriting historical findings.

Footguns#

Older product notes are prior research, not current compute evidence. RunComfy, ComfyDeploy and Floyo now have standardized refreshes linked above. Preserve the historical notes, but use current snapshots for tariff/API claims. Comfy Cloud's earlier profile predates this snapshot. ThinkDiffusion and Floyo share team/history; do not aggregate them as unrelated companies.

Civitai and Midjourney have older creator/monetization notes elsewhere; current standardized refreshes are linked above. Historical estimates and payout rules are not renewed as current facts.

Testing#

This combines public-source research with bounded authenticated diagnostics, not a matched performance benchmark. Comfy API catalog/node/system queries, two rejected URL-import requests and an approved private probe attempt were observed. Build creation returned HTTP 402; subsequent diagnostics confirmed funds but FREE tier, and a valid model-free Cloud workflow returned HTTP 403 FREE_TIER_NOT_ALLOWED. The original error is not proof of an empty balance. No GPU execution occurred. Callback services were stopped; removal of the uploaded 1.7 KB node blob is unconfirmed. No conversion experiment, audited finances or supplier contract was obtained. Source capture checks, citation review and local-link validation support reproducibility, not independent verification of marketing claims.

Final local checks covered eight manifests and 165 source receipts: JSON/ID validation, captured-byte counts, SHA-256, header-status consistency and referenced docket existence all passed. All 120 local Markdown links resolved; all 136 docket citation occurrences were reconciled to captures, canonical aliases or carried-forward supplier sources. The index also contains eight external discovery links. Secret-shaped credential/signed-URL checks and git diff --check passed for the authored documentation. Price arithmetic was independently recalculated. External citations were reviewed as evidence, not exhaustively availability-tested. No application runtime tests were needed for these documentation-only changes, and no successful GPU benchmark is claimed.

Operations#

Completed research lanes, retaining the original analyst prioritization:

1. Creator-community economics: PixAI, Leonardo and Mage.space. Investigate recurring free access, catalog/community distribution, paid retention and actual procurement.
2. Direct ComfyUI deployment alternatives: ViewComfy and Runflow, alongside refreshes of RunComfy, ComfyDeploy and Floyo. Separate hosted-user products, deployment services and customer-supplied infrastructure.
3. Hosted open-tool suites: ThinkDiffusion and RunDiffusion. Compare persistent sessions, subscriptions and creator acquisition with our job-based offering.
4. Broader creative-suite growth: Higgsfield, then compute/growth refreshes for Civitai and Midjourney. Their product scope and distribution may matter more than a matched ComfyUI workload.

All newly listed consumer/hosting candidates now have dated dockets. Leonardo and Midjourney direct pricing/documentation downloads were blocked (HTTP 403), despite readable web retrieval. Higgsfield live pricing was a shell; its exact checkout tariff remains unknown. Infrastructure findings rely on separate provider disclosures, not these incomplete captures.

Runware, fal and Replicate now have dated dockets in the separate inference/API lane above. They should not be mixed into a consumer-product growth ranking without matching the customer/use case.

For the next snapshot, capture sources before processing, retain publication and retrieval dates, add claim-level citations, record contradictions and update the manifest. Do not schedule automatic upkeep unless requested.

Continuation completion record — 2026-10-04#

The founder requested completion of the remaining research, closer Comfy infrastructure investigation and commits as verified batches finish. This checklist is the fixed continuation scope, not a new strategy decision. Unknown procurement or unavailable private metrics may be a documented conclusion; they must not be silently omitted or called verified ownership.

Every new batch must update coverage, source receipts and verification limits and be committed explicitly. A final synthesis must distinguish hardware ownership, controlled leases and usage-metered infrastructure; do not claim free access caused revenue growth without cohort/economic evidence.

All fixed-inventory public-source passes and the final synthesis are recorded. Further worker execution needs existing paid-plan access; procurement contracts, private cohorts and exact inaccessible checkout prices require new evidence, not indefinite re-reading of unchanged public pages. No automatic monitoring or paid upgrade was created.

Related#

• 2026-10-04 source manifest: new raw downloads, checksums, evidence roles and retrieval limits; local archives are not committed or uploaded.
• Comfy continuation manifest: new API product sources, live catalog observations, Runpod correlation and callback-test limits; raw authenticated records stay private.
• Creator-community batch manifest: PixAI, Leonardo and Mage sources, historical deployment dates and blocked-pricing archive limit.
• Hosting/app distribution batch manifest: ViewComfy, Runflow, ThinkDiffusion and RunDiffusion sources and product-boundary checks.
• Deployment refresh manifest: RunComfy, ComfyDeploy and Floyo current sources, commercial transitions and historical-offer corrections.
• Creative-suite refresh manifest: Higgsfield, Civitai and Midjourney sources, training/serving distinctions and capture limits.
• Creator-platform offer manifest: TensorArt, SeaArt and RunningHub creator rules, campaigns and inaccessible/contradictory offers.
• Adjacent API manifest: Runware, fal and Replicate procurement, conditional compute tariffs, developer incentives and company ownership.
• 2026-10-03 GPU-supply report and its archived source manifest: detailed TensorArt, SeaArt and RunningHub procurement evidence.
• Owned-GPU strategy: our operating boundaries, not competitor outcomes.